BCV-Predictor: A bug count vector predictor of a successive version of the software system

Abstract

Predicting the number of bugs in a software system intensifies the software quality and turns down the testing effort required in software development. It reduces the overall cost of software development. The evolution of hardware, platform, and user requirements leads to develop the next version of a software system. In this article, we formulate a problem and its novel solution, i.e., we are considering the prediction of the bug count vector of a successive version of a software system. After predicting the bug count vector in the next version of a software, the developer team leader can adequately allocate the developers in respective fault dense modules, in a more faulty dense module, more number of developers required. We have conducted our experiment over seven PROMISE repository datasets of different versions. We build metadata using a concatenation of different versions of the same software system for conducting experiments. We proposed a novel architecture using deep learning called BCV-Predictor. BCV-Predictor predicts the bug count vector of the next version software system; it is trained using metadata. To the best of our knowledge, no such work has been done in these aspects. We also address overfitting and class imbalance problem using random oversampling method and dropout regularization techniques. We conclude that BCV-Predictor is conducive to predicting the bug count vector of the next version of the software. We found five out of seven meta datasets reaches to more than 80% accuracy. In all seven meta datasets, Mean Squared Error (MSE) lies from 0.71 to 4.715, Mean Absolute Error (MAE) lies from 0.22 to 1.679, MSE and MAE over validation set lie between 0.84 to 4.865, and 0.22 to 1.709 respectively. We also compared the performance of BCV-Predictor with eleven baselines techniques and found the proposed approach outperform on most of the meta-datasets.